A verification harness lets coding agents earn trust

Cursor engineer Lauren Tan explains a progression from closely supervised agents to automated PR merges: agents must run the product, inspect evidence, and work inside explicit skills, feature maps, and enforceable constraints. Why it matters: More agents do not create leverage when every change still needs a person to discover whether it worked. A verification harness moves repeatable checks into the system before merge.

Try this: Inspect the last five recurring review comments in one repo, then turn one repeatable rule into a lint, CI check, or explicit architecture boundary and run it against a real agent-generated diff.